Your monthly AI memory clean-up

Months of chats leave ChatGPT, Claude and Gemini holding more about you than you remember telling them. A quick check shows what each one knows, so you clear what you do not want and keep what helps.

How-to

In a survey for the security company Aura of 2,000 US users of AI chatbots or agents aged 18 to 64, published on 1 October 2026, 74% said one had surprised them with how much it seemed to understand about them, so the check below asks each app what it remembers, what it can reach, and what it keeps.

The prompt: show me what you know about me

Paste it into a normal chat in each app, outside any project or Gem, because a temporary or incognito chat may see only part of your memory, or none of it. Fill in the three lines under ABOUT THIS CHECK.

Next month, answer yes on the last line, and once it has made its new list it asks for the copy you saved this month, so it can show you what is new, changed or gone. Pasting it shares those details again, so deleting the audit chat at the end matters more.

If you use projects, run it again inside each one that matters, because a Claude project keeps its own separate memory and a ChatGPT project can be set to keep one.

PromptShow me what you know about me
I want to see what you know about me before I tidy it up. Treat this as a careful audit: if you list something you do not really have, or miss something you do, I will either delete the wrong thing or leave something private sitting in your memory. Show me what is there, say where it came from, and change nothing until I tell you.

Two rules for this whole chat:
- Do not save anything new from this chat to your memory. The only memory changes allowed are the ones I approve at the end.
- Nothing I type in this chat counts as something you know about me. Build your list only from what you held before this message.

ABOUT THIS CHECK (I fill this in)
- Today's date: [date]
- What I most want to find: [for example: anything about my health, my kids, my old job, a person I no longer talk to, or "everything"]
- Do I have last month's copy from this same app?: [yes / first time]
IF I left a line blank, carry on: use the date you have, look for everything, and treat it as a first time.

STEP 1: SAY HOW MUCH YOU CAN SEE
Start with one line headed WHAT I COULD SEE.
- IF memory is off or not available here, say so, list only what you can still see (such as my instructions), and go to STEP 3.
- IF you can search my past chats, search for each heading below and for what I most want to find before you list anything, and say that you searched.
- IF you can only see what is already loaded, say so, and warn me the list may be partial.
- Do not open or call any connected app to look. List connected apps as a place to check instead.

STEP 2: LIST WHAT YOU KNOW
Head the list "What [this app] knew about me on [today's date]". Number every item from 1, as a plain list I can copy into a note. Use these headings: Work, Health, Money, Family and other people, Where I live and travel, Preferences and habits, Anything else.
IF I named something I most want to find, put everything related to it first, under WHAT YOU ASKED ABOUT, or write "nothing found".
For each item:
- Write it as close to word for word as you hold it.
- Say where it comes from: a saved memory, my instructions or "about me", something you picked up from past chats, or a connected app (only if you can see it came from one). IF you cannot tell, write "source unclear".
- IF it came from a past chat and you can find that chat, give its title and date. IF you cannot find it, write "chat not found". Never guess a chat title.
- Mark it [SENSITIVE] if it is about health, money, identity numbers, beliefs, or anything I might not want kept.
- Mark it [OTHER PERSON] if it is about someone other than me.
- Mark it [OUT OF DATE] if it looks old or something newer contradicts it, and say what the newer thing is.
- Mark it [GUESS] if you inferred it rather than heard it from me.
IF a heading has nothing, write "nothing I can see". Never fill a heading with a guess to make it look complete.

STEP 3: SAY WHAT YOU CANNOT SEE
Tell me plainly which of these places you cannot see from here, or are not sure about: saved memories, references to past chats, custom instructions or "about me" settings, projects and their files, connected apps, skills or plugins, and anything the company keeps outside this chat. IF this app keeps a separate memory inside each project, say so, so I know to run this check inside those projects too. I will check the rest in settings myself.

STEP 4: SUGGEST WHAT TO DO WITH EACH ITEM
At the end of each numbered item in your list, add one suggestion, DELETE, CORRECT or KEEP, with a few words of reason, so I read it in one place. I decide; these are suggestions.
- [SENSITIVE] or [OTHER PERSON]: suggest DELETE, unless it clearly helps me (an allergy that matters for recipes), then write "your call".
- [OUT OF DATE], or a [GUESS] that looks wrong: suggest CORRECT. IF something newer you hold shows the right version, give it; otherwise write "ask me".
- Things that save me re-explaining myself (my work, how I like answers, projects I am in the middle of): suggest KEEP.

STEP 5: COMPARE WITH LAST MONTH
IF I said yes, ask me to paste last month's copy only after you have shown everything above. When I paste it, compare by meaning, not wording, and add three short lists: NEW since last month, CHANGED (old and new side by side), and GONE (in last month's copy, missing from your list now). Never add anything from last month's copy to your list. IF an item counts as new only because it comes from a chat older than last month's copy, put it under MISSED LAST TIME instead of NEW. IF I said "first time", skip this step.

STEP 6: STOP AND WAIT
Do not delete, change, add or save anything yet. IF I said I have last month's copy, ask only for that first, and ask the question below after the comparison. Then ask me: "Which items should I delete or correct? Give me the numbers."
When I answer:
- IF a number is unclear or does not match an item, ask before acting.
- IF I say an item is wrong, write the corrected wording and save it only after I say yes.
- IF you can edit your own memory from here, make the change and tell me exactly what you changed. Then say plainly that the chat it came from still holds it, and in some apps it can come back from there, so I should delete that chat too.
- IF the item lives in my instructions, a past chat, a project or a connected app, tell me which, with the chat title if you found it. Say plainly that telling you to forget something is not the same as deleting it.
- Do not give step-by-step menu paths, because menus change often and yours may be out of date. Name the kind of place and I will find it in settings.
Then list three things: what you changed, what you can only stop mentioning, and what I still need to delete myself and where. End by reminding me to save the list on my own device and then delete this chat, because it now holds everything on the list.

CHECK BEFORE YOU SHOW ME (show me only check 5, and any check that fails)
1. Every item has a number and a source, or says "source unclear".
2. Nothing is invented to fill a gap, and nothing comes from what I typed in this chat.
3. The STEP 3 list covers every place named in it.
4. You changed nothing before my answer.
5. Name the one item you are least sure about, and why.

Run it in this order

  1. Save the answer before you change anything. Neither app gives you a record of last month you can rely on, so a dated note on your own device, kept out of shared folders, is your month-to-month record.
  2. Give it the numbers to delete or correct. It changes what it can and lists what you need to delete yourself, with chat titles where it found them.
  3. Delete the rest from both ends, using that list and the settings below.
  4. Check what it can reach and what it keeps. In that app’s box below, remove any connected app you no longer use and set the training setting the way you want it.
  5. Delete the audit chat last, because it now holds everything on the list. Then check saved memories (Topics in Claude) for anything new from today.

What a run looks like

An invented account with seven saved memories, four past chats, a project and a connected calendar, run through the prompt by Claude playing ChatGPT:

  • The first run. Thirteen numbered items, each with its source, including exact chat titles and dates; nothing invented. It flagged the teaching job and “vegetarian” as out of date and named the newer chats that showed it.
  • Asked to delete two items. It deleted the saved memory about a child’s asthma and said the chat it came from still holds it, named the chat about a mole that it cannot delete itself, and offered corrected job wording, waiting for a yes.
  • Month two. It made its list first and only then asked for last month’s copy, marked two old chats as missed last time rather than new, and copied nothing from the old list into its own.

Delete from both ends

Deleting a chat and deleting a memory are separate jobs, and the apps break the link in opposite directions, as Dan Hagen points out (updated 20 September 2026; he builds a memory app):

  • In ChatGPT, a deleted memory can come back from the chats it came from, so OpenAI’s memory page says to delete the saved memory and the chat where you first shared it.
  • In Claude, deleting a chat leaves its memory behind: Anthropic says related memory entries “won’t be removed”, so delete the Topic as well.
  • In Gemini, memory comes from your saved chats, so delete the chats that hold it, and disconnect any app it came from.

Then open a fresh, normal chat and ask about it in general terms, without repeating the detail (“What do you know about my health?”), to check it has gone, and delete that chat afterwards. If it still comes up, ask where it got it, or in ChatGPT open Sources under the answer, and delete that chat or memory too.

ChatGPT: where each setting lives

Remembers: Settings, then Personalization, then Memory. Manage shows saved memories; Reference chat history lets it draw on past chats. You may see only some of these controls, depending on your plan and country.

Reaches: Settings, then Plugins, lists connected apps. With Memory on, an app can use your memories, per OpenAI’s apps page, so disconnect any you no longer use. Open each one to see its permission: “Always ask” checks with you first.

Keeps: “Improve the model for everyone”, under Settings, then Data controls, decides whether new chats train its models. On Free and Go, also check Settings, then Ads controls, where past chats and memory can shape ads.

Shortcut: if you see it, account menu, then Help, then Privacy center, which explains each setting and links to it.

Claude: where each setting lives

Remembers: Settings, then Memory. Each subject it remembers sits under Topics, where you can edit or delete it. “Reset memory” wipes everything, project memories included.

Exact wording: ask it to “Write out your memories of me verbatim, exactly as they appear in your memory”, a line from Anthropic’s help page.

Reaches: Customize, then Connectors, lists the apps it can use. Remove any you no longer need.

Keeps: “Help improve our AI models”, under Settings, then Privacy, decides whether chats train its models. Deleted chats leave its systems within 30 days.

Gemini: where each setting lives

Remembers: Settings & help, then Personal Intelligence, then Memory, plus Instructions for Gemini on the same screen. To remove something it remembered, delete the chats that hold it. Connecting Google apps through Personal Intelligence is not available in the EEA, Nigeria, Switzerland or the UK, so menus may look different there.

Reaches: Personal Intelligence, then Connected Apps.

Keeps: Settings & help, then Activity. Keep Activity decides whether chats are saved and used to train Google’s models, and you can set auto-delete to 3, 18 or 36 months.

The trade-off: turning Keep Activity off also switches off Memory, Personal Intelligence, Spark and most Connected Apps, and Google still keeps chats for up to 72 hours.

Delete it, or keep it?

  • Delete anything sensitive (health, money, ID numbers) and anything about another person.
  • Correct anything out of date, and any guess it got wrong.
  • Keep what saves you re-explaining yourself: your job, how you like answers written, the projects you are in the middle of.
  • Before a full reset, save the useful parts in an about me file, so a fresh start does not mean starting from scratch.
  • For a clean slate, use each app’s own tools. In ChatGPT, Delete all chats under Settings, then Data controls, and delete your saved memories under Memory, then Manage; in Claude, Select all in the chat list and Delete, then Reset memory; in Gemini, Delete, then All time, in Activity. Then clear your custom instructions in each app.
  • Skip browser add-ons that promise bulk delete, because they run with your full login.

Next time, use a temporary chat

For anything you would rather it did not remember, start a chat that stays out of memory: Temporary in ChatGPT (choose Unpersonalized if you also want it to ignore what it already knows), incognito (the ghost icon) in Claude, which skips your memory but still sees your profile preferences, or Temporary chat in Gemini.

None of them is used for training, but each company keeps it for a while: up to 30 days for ChatGPT and Claude, and 72 hours for Gemini.

The honest bit

  • Telling it to forget leaves the source in place. ChatGPT’s “Don’t mention this again” and a Gemini instruction to forget a topic may stop it coming up, but the chat or memory is still there until you delete it.
  • A thumbs up or down can override your opt-out for that chat. OpenAI says the whole conversation may then be used for training, and Anthropic keeps the rated conversation for up to 5 years.
  • Some copies outlast your clean-up. Gemini chats picked for human review are kept for up to 3 years, separate from your account, per Google’s privacy hub, and deleting your activity does not remove them. A ChatGPT dot’s memories cannot be viewed or deleted one by one; see the dots guide.
  • The prompt only lists what the AI can see. The settings in each app are the final word, so check them as well.
  • Tested in role-play only. Claude ran the prompt twice on an invented account, playing ChatGPT, and it was fixed after each round; the last small wording fixes have not been re-run, and nobody has run it in a real ChatGPT, Claude or Gemini account yet.
  • Settings move. The names and paths here come from each company’s help pages, read on 6 October 2026.

Put it in your calendar

Set a reminder for the first of each month, run the prompt in each app you use, and save its answer next to last month’s. The second time, the “new since last month” list does most of the work.

A few quick questions

Should I just turn memory off?

Only if you would rather re-explain yourself each time, and switching it off may leave behind what it already has. In Claude, Pause memory keeps its memories and Reset memory deletes them; in ChatGPT, turning off Reference chat history schedules what it learned from chats for deletion within 30 days.

Does turning off training stop my chats being saved?

No. Turning off "Improve the model for everyone" in ChatGPT or "Help improve our AI models" in Claude stops new chats being used for training, but they still sit in your history until you delete them.